NVIDIA opens its frontier self-driving model for commercial use

NVIDIA Alpamayo 2 Super is a frontier open reasoning model built specifically for robotaxis and autonomous driving development. Built on NVIDIA Cosmos 3 Super Reasoner and post-trained with reinforcement learning, it is now available on Hugging Face under the permissive OpenMDW-1.1 license for commercial deployment.

The model is three times the scale of the earlier 10-billion-parameter Alpamayo 1 and 1.5 models. By adopting the Linux Foundation's OpenMDW-1.1 license across the entire model family, NVIDIA gives developers, automakers, and suppliers the legal right to fine-tune, build derivatives, and deploy commercial fleets without proprietary API lock-in. The Alpamayo family has already surpassed 500,000 downloads on Hugging Face.

What it can do

Alpamayo 2 Super tackles the rare, dangerous edge cases that conventional vision pipelines fail to anticipate. It fuses full 360-degree surround camera inputs from the front, sides, and rear of the vehicle, allowing it to navigate complex intersections, unprotected turns, and crowded merges.

For any driving situation, the model produces five tightly integrated outputs: a trajectory detailing the vehicle's planned path, a chain-of-causation (CoC) trace explaining why it chose that maneuver, a high-level meta-action such as yielding or changing lanes, reasoning auto-labels for downstream datasets, and visual question answering grounded directly to 2D image coordinates.

This makes self-driving decisions fully inspectable rather than a black box. The chain-of-causation traces integrate directly into NVIDIA Halos safety-validation workflows and meet ISO/PAS 8800 safety engineering standards.

On the LingoQA autonomous driving reasoning benchmark, Alpamayo 2 Super ranks first among nearly 40 evaluated systems. Using the Lingo-Judge metric, it beats Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1 points, and GPT-4o by 23.2 points. It also swept first place across all autonomous driving benchmarks evaluated by NVIDIA.

How to deploy it

Rather than running raw on vehicle hardware, Alpamayo 2 Super is engineered for a cloud-to-car workflow. Teams run the heavy model in cloud infrastructure to produce high-grade reasoning traces, synthetic training scenarios, and teacher labels, which are then distilled into compact models for real-time inference inside the vehicle.

It also functions as an automated fleet annotator. By linking causal reasoning to 2D image regions, developers can convert raw camera clips into rich training data in days instead of spending months on manual labeling.

The broader stack includes the AlpaSim closed-loop simulator, the AlpaGym high-throughput reinforcement learning framework, and NVIDIA Physical AI Open Datasets. Weights, autolabeling pipelines, and training recipes are live on Hugging Face right now.

Why it matters

For autonomy startups and carmakers, this gives you a world-class reasoning foundation without signing away your fleet data to closed API providers. Because OpenMDW-1.1 allows unrestricted commercial derivatives, engineering teams can build custom driving policies entirely on private infrastructure.

Unlike general multimodal models like GPT-4o, Alpamayo 2 Super outputs native path coordinates and causality links tailored to vehicle kinematics. If you run a data pipeline for AVs or robotaxis, deploying this as your cloud-side teacher model looks like an easy win.